Acquisition Time Reduction in Pediatric PET Imaging using Unsupervised Convolution Neural Networks

The deep image prior (DIP) method, based on unsupervised training of convolutional neural networks, has demonstrated the potential of using neural networks as a constraint in image denoising. An appropriate stop criterion for the DIP method is crucial to avoiding insufficient training or overfitting. We propose a DIP-based image denoising method for PET images corresponding to 1/3 of the clinical standard 90 sec/bed acquisition time using auto-stop criteria. PET/CT data from 7 patients under age 12 were used in this study. For each patient, we trained a U-Net with his/her whole-body CT image as the input and the 30 sec/bed PET image reconstructed by the TOF Q.Clear with a regularization parameter (RP) of 700 as the label. To properly train the U-Net, we adopted self-validation based on the low-frequency information of the input and the label. The training iteration was stopped automatically when the validation loss began to increase or converge. We compared the DIP outputs with the Gaussian filtered 30 sec/bed images (RP = 700) and the 30 sec/bed images reconstructed with the RP of 1300, with the clinical 90 sec/bed images (RP = 700) as references. The DIP-generated images displayed lower noise and were visually closer to the reference images. The average normalized mean squared error values (×10
^-5
) were 5.9 ± 1.5 for the DIP-generated images, 17.6 ± 3.0 for the 30 sec/bed (RP = 700) images, 10.6 ± 1.9 for the Gaussian filtered images, and 8.7 ± 1.2 for the 30 sec/bed (RP = 1300) images. With the 90 sec/bed images as references, the average SUV
mean
, SUV
max
, and SUV
standard

deviation

(SD)
/SUV
mean
differences decreased by 10%, 84%, and 80% after the DIP enhancement. Quantitative results showed that the DIP method with auto-stop criteria improved the image similarity and quantitative accuracy of images from shortened acquisition time with respect to the clinical standard images, demonstrating its potential in reducing acquisition time for pediatric PET imaging.